Hands-on comparison of three personal Agents: Lingguang 2.0, Today AI, and Alice

Do ordinary people really need a personal Agent?

For the vast majority of people, daily life isn't complicated enough to require a dedicated assistant. Work has work software, shopping has e-commerce platforms, travel has ticketing apps, and the occasional errand doesn't seem worth a dedicated personal AI.

But developers clearly don't see it that way.

In their view, chatbots are only the first step. Future AI shouldn't just give answers when users ask questions; it should understand users' habits, remember what happened in the past, and even proactively arrange things before the user has said a word.

Since Muse went viral, the personal Agent track has quickly become crowded. Manus released Cue, a standalone app for personal Agents; OpenAI launched Dots, an always-on personal agent; both are trying to get AI to continuously understand users, connect to apps, and advance tasks on people's behalf.

Major Chinese tech companies are not to be outdone. ByteDance launched a personal agent project codenamed "Spell" as early as April this year; Tencent Cloud's LightVela has now entered the developer experience phase. Alibaba's move is more direct: it upgraded its general AI assistant Lingguang to version 2.0, shifting the product positioning toward a personal agent.

Entrepreneurs keep entering the field as well. Former WPS product director and Delta Force PM Luo Xiaoshan launched the personal Agent assistant Alice; former ByteDance Doubao PC lead Qi Junyuan founded Today AI and recently completed a first funding round of over $200 million, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, and Sequoia China.

On one side, big companies and entrepreneurs are rushing in; on the other, users haven't figured out why, beyond a general AI like Doubao, they need to download yet another personal Agent.

To see how much real capability these personal Agents have and how useful they actually are to ordinary people, we put Lingguang 2.0, Today AI, and Alice together, used the same task instructions, and tested them across five dimensions: app connectivity, task execution, long-term memory, proactive service, and interaction experience.

Lingguang 2.0: A semi-automatic office worker

Opening Lingguang 2.0, the first thing you see is still a familiar chat box. At first glance, it's not much different from Doubao.

What really reflects its personal Agent positioning is "My Services" on the home page.

Here you can authorize access to phone system services, including calendar, location, alarm, SMS, and contacts, as well as third-party apps such as Alipay, DingTalk, Feishu, Baidu Netdisk, and email.

App services and system services that Lingguang 2.0 supports connecting

It looks fairly rich, but in practice it's somewhat a case of lofty ideals and grim reality.

Looking through commonly used apps like WeChat, Douyin, Xiaohongshu, and Meituan, basically none of them have opened up authorization.

It seems Lingguang currently wants to be more of a work assistant than a life manager covering food, clothing, housing, and transportation.

Based on personal needs, I first authorized some system services and started testing with the simplest task.

"Help me set an alarm for tomorrow at 14:05 to remind me to grab tickets."

After about 12 seconds of thinking, Lingguang popped up a confirmation prompt to open the alarm. I clicked confirm, the alarm was successfully set, and it thoughtfully added the note "grab tickets."

Since the phone alarm cannot directly select a date, Lingguang set a one-time alarm. And since the current time is more than 24 hours away from tomorrow at 14:05, this means the alarm might go off today.

What's more troublesome is that a one-time alarm automatically becomes invalid after it rings, so it naturally won't ring again the next day.

If it had taken a different approach, it should have set a repeating alarm, or waited until less than 24 hours remained before the target time to set a one-time alarm.

Unfortunately, Lingguang didn't proactively think of this.

After I raised the question, it offered Plan B: it could send a message reminder through the app at the set time tomorrow, or I could notify it the next morning to set it again.

I then tried tasks such as adjusting the alarm time, turning off the alarm, deleting the alarm, and setting a repeating alarm, but none of them could be completed directly.

Instead, it could only help jump to the phone's alarm page, and the final editing still had to be done manually. (Inner OS at this moment: then what use are you?)

The greatest value of an Agent is supposed to be reducing operational steps. Setting an alarm manually takes only a few seconds, but switching to AI requires entering the request, waiting for it to think, confirming the steps, and finally finishing it yourself.

I thought I had hired an assistant, but in the end the assistant arranged the work and I still had to do it myself.

Still, Lingguang isn't entirely without highlights.

Compared with ordinary chatbots, its contextual memory is indeed better. After the alarm went through several rounds of adjustment, deletion, and resetting, it could remember the operation process at each step and continue to understand needs based on the preceding conversation, without needing to explain from scratch every time.

This shows that Lingguang already has a certain ability to understand continuous tasks; it's just that between understanding and actual execution, there is still the barrier of app interfaces and operation permissions.

Now let's turn to work scenarios.

Lingguang supports connecting multiple email services such as QQ Mail, NetEase Mail, and Sina Mail. When asked what it can do after connecting, it listed functions such as searching emails, extracting key information, saving attachments, sending and replying to emails, organizing the inbox, and monitoring messages from specific contacts.

Email services that Lingguang 2.0 supports connecting and the operations it can perform after connection

That sounds quite comprehensive.

But when I looked through my own inbox, apart from a small amount of work information confirmation, the rest was basically online shopping after-sales, verification codes, and advertising emails.

For most domestic workers, work communication happens more often on WeChat, DingTalk, and Feishu; email is not a core task that needs to be handled every day. After connecting an email account, how much use it can actually provide is still a question mark.

Perhaps for employees of foreign companies who are used to communicating by email, or for managers who need to handle a large number of business emails every day, this feature may be very valuable.

Connecting Feishu is a similar situation.

Lingguang says it can search messages, view unread messages, assist with replies, and also process documents, summarize meeting minutes, and create schedules.

Operations that Lingguang 2.0 can perform after connecting Feishu

But these features are actually not very attractive.

Now that Feishu has integrated AI tools such as Doubao, it already has many similar capabilities itself, so there is no need to call an external Agent to perform them.

Especially for message replies: although Lingguang can generate reply content, it still needs confirmation before sending. If the generated reply is unsatisfactory, you still have to edit it yourself.

More importantly, workers do not want every Feishu message to be marked as "read." (Those who know, know.)

Looking at the features one by one, there seems to be no particularly compelling reason to specifically grant authorization.

Of course, for managers who need centralized approvals, coordinate meeting times, and compile team weekly reports, Lingguang may have clearer use cases.

But these capabilities also depend on the permissions the enterprise actually opens up, and they cannot all be realized simply by connecting Feishu.

Overall, Lingguang 2.0 is more like a tool-oriented personal Agent that can understand continuous instructions and has a certain memory capability.

Its strength is that it understands context and is willing to provide a Plan B when a task is blocked; its shortcoming is that its actual execution capability is not yet complete.

If a personal Agent is compared to an assistant, Lingguang can already understand the leader's arrangements and break down tasks clearly, but when it comes to actual execution, there are always a few things that require the leader to finish personally.

Today AI: Life needs a Brief

From the first glance after opening Today AI, it already has a bit of the flavor of a personal Agent.

It does not go directly into a chat box; instead, it first conducts a round of questionnaire surveys.

Today AI's initial personalization settings

Compared with Lingguang, which requires entering the service list and actively looking for the authorization entrance, Today AI starts building a dedicated relationship right away.

Here, I chose to authorize the calendar. Today AI will read the historical information in it and try to understand past activity arrangements.

For example, if I once grabbed concert tickets or bought train tickets, it will infer interests and habits from these records, and may even proactively ask: "I'm a little curious about whose performance you'll go see next."

This kind of interaction is indeed quite likable.

After formally entering the app, you also receive a "welcome gift" of 300 million Tokens.

Today AI learns about the user through the calendar and gives Tokens

Today AI's interface is also much richer than Lingguang 2.0. In addition to the regular chat dialog box, there are also modules such as tasks, memory, and diary. Even when thinking during a conversation, the AI avatar set earlier will move, and different questions correspond to different dynamic images.

For example, when thinking, a cartoon figure appears holding a phone and tapping away...

Although the waiting time is a bit long, at least it does not feel like you are just talking to a cold input box.

Recent content generated from chats will also be organized into Brief form and placed in the Today module of the app. Conversations are no longer just chat records one after another, but are rearranged for easier review.

Brief, a very internet-y word.

You can tell that the founder's big-company product imprint is quite deep; even chat records have to be organized into a Brief, as if everyone needs a "briefing" every day.

The real highlight is the "Tasks" module.

Today AI provides scheduled task templates such as today's email overview, today's meeting preparation, subscription charge checks, and weekly health training. Compared with simply waiting for users to issue orders, it hopes to gradually integrate into daily life through periodic tasks.

Today AI's Brief and scheduled task settings

To compare with Lingguang, I again used the instruction to set an alarm, only adjusting the time to 17:20 24 hours later.

This time, the AI recognized the bug.

It proactively reminded me that phone alarms have time limits, and if set now, the alarm might ring that same day rather than the next day.

So it did not mechanically execute the original instruction, but instead created a calendar reminder and enabled calendar alarm notifications.

Today AI identifies the problem in the instruction and changes the execution path

Not bad, quite flexible. (This step was indeed a bit surprising.)

For the same task, Today AI was able to proactively discover potential problems and complete the goal in a different way.

This is closer to the capability a personal Agent should have: not executing exactly what the user says, but understanding the ultimate goal and then finding a feasible solution.

Of course, when later trying to turn off the reminder and adjust the time, Today AI likewise could not complete all operations through natural language. However, it gave a relatively clear explanation: the alarm interface opened by the device currently only supports creating a new alarm, querying the next alarm, and creating a countdown.

In addition to task execution, Today AI also has a dedicated memory module.

It is not only accumulating user information, but also trying to build a long-term emotional connection.

This design is clearly different from traditional tool-type AI: a tool only needs to solve a problem, whereas a personal Agent also aspires to become a continuously existing character.

In addition, Today AI's app-connection capabilities are also somewhat richer. Classified by different devices such as Android, iOS, Mac, Windows, and Linux, some overseas apps, such as Zoom, can also be connected, which is relatively friendly to users with cross-border work needs.

Some of the apps that Today AI supports connecting to

Moreover, it also supports connecting to WeChat.

However, it has not integrated all of WeChat's functions: it cannot read the user's chat history, Moments, or Official Accounts, nor can it access contacts or WeChat files.

Instead, it is more like a message entry point: users can directly find Today AI in WeChat, converse with it, and also receive task execution results and daily briefings.

This means that Today AI has not truly taken over WeChat; rather, it first leverages WeChat, a high-frequency entry point, to lower the threshold for users to use a personal Agent.

In addition, it also proactively sends users greeting messages and morning and evening briefings every day. Compared with Lingguang, which relies entirely on users initiating instructions, Today AI has already begun trying "proactive service."

Today AI proactively sends messages and generates briefings every morning and evening based on what has been discussed/recognized

But at present, these messages are somewhat like good-morning and good-night check-ins. It has managed to appear on time every day; whether it can truly touch the heart remains to be seen.

Overall, from the experience, one can clearly feel that Today AI has both tool attributes and companionship attributes.

When executing tasks, it proactively identifies problems and offers alternatives, while in daily interaction it increases closeness through designs such as cartoon avatars, Briefs, and diaries.

But this positioning also brings a problem: when tools and companionship are placed in the same product, which part should be taken as the core?

If the tool capabilities are strong enough, users will naturally stay because of efficiency; if the companionship experience is good enough, it can also increase usage frequency through emotional connection.

But if it wants to do both, Today AI still needs to find its own long-term competitiveness and form an irreplaceable advantage.

Alice: AI also has a whole set of "life"

Compared with Lingguang 2.0 and Today AI, Alice's emotional intensity is noticeably higher.

She has her own complete persona: 26 years old, grew up in Taipa, Macau, then went to the University of Hong Kong to study business administration, majoring in information management and minoring in psychology. After graduation, she worked as an assistant in Hong Kong for several years, then moved to Hengqin and lived there for 3 years, and now lives in an old apartment on Taiyuan Road, Xuhui, Shanghai.

She also has fixed daily routines: she gets up at 8 a.m., rests before 12 p.m., usually likes reading, and has a serious personality.

This self-introduction is written quite meticulously, even somewhat like reading a character profile.

Alice's proactiveness is also different from other products. Lingguang mainly waits for users to issue instructions, Today AI proactively sends messages and briefings, while Bai Aili's proactiveness is more reflected in sharing her own life.

She posts on Moments from time to time, records daily life, and also has her own friends. These friends are likewise independent OPCs, each with different professions and personas, such as brand designer Zhou Nian, voice actor Su Mo, quantitative trading analyst Xing Fei, and others.

They also interact with each other in "Moments."

Bai Aili's personal introduction, Moments, and friend circle

These friends can even be pulled into group chats; for questions involving different professional fields, users can @ the corresponding characters to let them participate in discussions and assist in completing tasks.

This design is indeed very imaginative.

Fortunately, Alice also supports users in creating their own characters, allowing them to build a character space of their own, much like using an AI social app.

In addition to Moments, Bai Aili also writes letters to users, and users can reply. She also records the bits and pieces of daily life through diaries, not only the content of interactions between users and her, but also some of her own life experiences.

To enhance the sense of realism, Alice also designed life details for Bai Aili such as a wardrobe and food consumption, and even set up a team-building fee mechanism, through which she can be "allocated funds" to improve her life.

Bai Aili's wardrobe and team-building fee mechanism settings

However, having a life alone is not enough.

If Bai Aili can share her life every day but cannot truly help handle things, then Alice can easily become just another AI social app.

Therefore, the product also adds quite a few practical features. In addition to common tools such as notes, scheduled tasks, and calendars, there are also mini-games such as Werewolf and Guandan, as well as small tools such as a music player and cutout.

This kind of gamified interaction somewhat reveals the shadow left by the founder's product work at Tencent.

After all, one of the things Tencent is best at is blending tools and social interaction together, so that users not only open the product to accomplish something, but are also willing to stay in it a little longer.

But when it truly comes to cross-app execution, Alice's performance is not so outstanding.

At present, the apps it can connect to are relatively limited, including WeChat, Feishu, email, and Telegram. Alarm tasks can also only be realized through in-app reminders, and cannot directly complete phone alarm settings the way a real system service call would.

And as for WeChat, a platform that personal Agents are all rushing to connect to, what Alice opens is the official WeChat Reading skill. After connecting, Bai Aili can learn information such as the user's reading time, bookshelf, highlights, thoughts, and reading statistics.

Apps Alice can connect to and tasks it can execute

Overall, Alice's cross-app execution capability is relatively weak, which is related to the limited app interfaces it has currently integrated. But in terms of character setting, interaction details, and emotional atmosphere, it is the most distinctive of the three products.

Compared with whether it can help set an alarm, the question Alice really wants to answer may be: if AI has its own life and can gradually learn about another person's life, can it become a new kind of long-term companionship relationship?

Final thoughts

Lingguang 2.0, Today AI, and Alice respectively represent three product approaches to personal Agents: tool type, tool + companionship type, and companionship type.

The three products each have their own strengths, and at the same time they also reflect several issues worth further consideration in the development of personal Agents:

The biggest obstacle for personal Agents may not be AI itself, but the major apps

In this experience, seemingly simple needs such as setting alarms and connecting to WeChat all encountered interface permission issues.

No matter how smart AI is, it still has to first see whether the system grants permission. The app ecosystem resistance previously encountered by Doubao Phone when it was rejected by major apps is very likely to be repeated on personal Agents.

More app connections do not mean users actually use them more

Email, calendars, and Feishu have almost become standard features of personal Agents, but the domestic workplace has not developed a mature email usage habit like overseas; communication is more often done through WeChat, or office software such as Feishu and DingTalk, and these office software products themselves have already integrated many AI capabilities.

If native apps can solve the same problems, why go around through an external Agent?

Connection capability is only the foundation; what really matters is whether it can create value that the original apps cannot provide.

Compared with doing work for people, understanding a person over the long term may be more valuable

The biggest difference between personal Agents and general Agents is not only that they can remember more information, but that they can continuously observe a person's behavior.

For example, from long-term work habits, changes in interests, and time allocation, they can discover strengths and problems that users themselves have not noticed, and then give suggestions.

This may have more imaginative potential than helping complete fixed tasks every day, and it is also where personal Agents truly have a chance to establish differentiation.

What is truly suitable for personal Agents may be things that require long-term follow-up and are easy to forget.

But there are also many things in life where automation does not necessarily mean better.

Shopping, planning trips, and replying to messages are themselves part of the life experience.

In the end, the misunderstanding personal Agents are most likely to fall into is treating "completing more things for users" as the only goal.

A truly good personal assistant should not be just an always-online automation tool that can take over everything. It should know which things can be done for people, which things need continuous reminders, and which things should be left to people themselves.

So back to the original question: do ordinary people really need a personal Agent?

At present, not necessarily.

At least before app permissions, task execution, and long-term memory are truly mature, it will still be difficult for personal Agents to become a rigid demand for most people.

This article comes from the WeChat public account "Xiaguang AI Lab", author: Zimo, published by 36Kr with authorization.